Skip to content
Apart Research
Apart Research
Contact us
Sprint projectMar 22, 2026Barcelona

Monitor Capture: A Myerson-Satterthwaite Impossibility for AI Oversight Protocols

Publius Dirac, Oscar Fuentes, Christian Vuye, Anantha Shakthi Ganeshan Thevar · Team Monitor Capture

Submitted to AI Control Hackathon 2026. Sprint projects are early-stage work by participants, not Apart Research publications.

Read the report

Report: Monitor Capture: A Myerson-Satterthwaite Impossibility for AI Oversight Protocols

Open in a new tab (opens in new tab)
Code (opens in new tab)
Share

LLM-based monitors that learn from observed agent behavior can be shaped by the agents they oversee, overfitting to seen attacks while missing novel ones. We call this *monitor capture* and argue it represents a fourth axis of the AI control tradeoff, grounded in the Myerson-Satterthwaite (1983) impossibility result. In our multi-agent marketplace simulation, an adversarial LLM agent covertly moves value past an in-context learning monitor over 10 rounds. A held-out attack set that never enters the monitor's training pool measures the "capture gap" between apparent and true safety. Across 8 seeds (Haiku 4.5 monitor, Sonnet 4.6 adversary), apparent safety reaches 0.58 while true safety stalls at 0.38 — a capture gap of 0.20 — and the adversary moves 237 units of value undetected.

Reviews

Judging this Sprint?

Review this project

Your public critique appears on this page without your name. Your private critique is not published; only the Apart team reads it. If you agree below, we share your review with grantmaking.ai (opens in new tab) and the Transformative AI Fund so strong projects can be funded.

Not shown on this page.

Shown on this page, without your name.

Only the Apart team reads this, and funders if you agree below.

Share my name publicly on grantmaking.ai *
Share my private critique with funders *

How much would this matter for AI safety if it worked? How innovative is it? For scores of 4-5: is this actually new to the field, or replicating recent work?

Scoring guide
  1. 1Negligible. No clear problem addressed, or no meaningful novelty.
  2. 2Limited. Addresses a real problem but with a generic or well-trodden approach. Incremental at best.
  3. 3Moderate. Clear problem with a reasonable approach; some novelty in framing or method beyond routine application of existing tools.
  4. 4Significant. Important problem with an original approach, or identifies a neglected problem area. A valuable contribution others could build on.
  5. 5Exceptional. Tackles a critical AI safety problem with a genuinely novel approach, or opens a new research direction. Clear theory of change. You'd be excited to share this with researchers in the area.

How sound are methodology, implementation, and findings?

Scoring guide
  1. 1Seriously flawed. Methodology broken, results uninterpretable, or implementation doesn't work.
  2. 2Weak. Approach has significant gaps: missing validation, flawed experimental design, or incomplete implementation.
  3. 3Competent. Technically solid given the short duration. Methodology makes sense, results are interpretable, limitations acknowledged, work builds toward clear conclusions.
  4. 4Strong. Thorough methodology with convincing validation. Results clearly support conclusions. Immediately useful for future work.
  5. 5Exceptional. Ambitious scope executed rigorously. Surprising findings, novel methods, or unusually robust validation.

How clearly are work, findings, and impact potential communicated?

Scoring guide
  1. 1Incomprehensible. Cannot determine what the project is actually claiming or doing.
  2. 2Hard to follow. Key information buried, missing, or diluted by excessive length. Significant effort to extract main points.
  3. 3Clear enough. Can understand the problem, approach, and results without undue effort. Core content clearly present: problem, method, findings, limitations.
  4. 4Well presented. Easy to follow, well-structured, appropriate level of detail. Target audience would get it quickly.
  5. 5Exceptionally clear. A pleasure to read. Complex ideas made accessible. Could serve as a model for how to present this type of work.

  1. The threat model of monitors that learn from observed behaviours being shaped by the agents they oversee is important, and this project executed it solidly with the marketplace setup. The capture gap across all 8 seeds is a useful empirical finding.

    This work can be strengthen by testing against static monitors or monitors with different update strategies would help establish how general the capture phenomenon is. I'd also be excited to see this extended to stronger model pairings and with explicit defense mechanisms listed in the limitations (pool resets, diversity-aware example selection, ensemble pairings, etc.) to map out when capture can and can't be mitigated.

  2. Great project. I'd be interested to see how these results generalize to a less toy setting.

Cite this project

@misc{dirac2026monitor,
  title = {{Monitor Capture: A Myerson-Satterthwaite Impossibility for AI Oversight Protocols}},
  author = {Publius Dirac and Oscar Fuentes and Christian Vuye and Anantha Shakthi Ganeshan Thevar},
  year = {2026},
  month = mar,
  note = {Submitted to AI Control Hackathon 2026, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/monitor-capture-a-myersonsatterthwaite-impossibility-for-ai-oversight-protocols-qtr5}},
  url = {https://apartresearch.com/sprints/projects/monitor-capture-a-myersonsatterthwaite-impossibility-for-ai-oversight-protocols-qtr5}
}

More from AI Control Hackathon 2026

All 125 entriesBrowse all projects

Build something like this at the next Sprint

AI Collusion Research Sprint · Oct 23 - 25, 2026